Linking clinical text to standardised SNOMED CT concepts is a key step for making electronic health records machine-readable, but it is difficult when labelled training data are scarce. A new arXiv preprint describes SNO-PI, which the abstract reports as the top-performing submission on SNOMED International's entity linking benchmark as of 18 October 2026.
On the benchmark, SNO-PI achieved a macro character IoU of 0.5387, outperforming the organisers' reference system. The paper's title indicates that the approach relies on pre-trained encoders combined with curated terminology knowledge to compensate for the lack of labels.
The paper is a preprint and has not been peer-reviewed. The abstract provides only high-level results, so details about the architecture and evaluation are not yet available from this summary. Still, the result suggests that combining general-purpose encoders with domain-specific terminology resources can be effective for clinical entity linking.